A lightweight Rust MCP server that searches Slack, Confluence, JIRA, GitHub, and local files in parallel from a single tool call.
Why
Developer knowledge is scattered across Slack threads, Confluence pages, JIRA tickets, and local codebases. Finding "what did we decide about X?" means manually searching 4+ systems.
Unified search MCP server — search Slack, Confluence, JIRA, and local files in parallel
ดาว
1
เครื่องมือ
—
อัปไทม์
—
This MCP server gives your AI assistant one unified_search tool that fans out queries to all sources in parallel, merges results with configurable ranking, and returns a single ranked list with comments included.
Features
Parallel fan-out -- all sources queried simultaneously via tokio
Cross-source ranking -- weighted relevance scoring with deduplication
Comments included -- JIRA and Confluence search results automatically include recent comments
Deep-dive lookups -- get_detail fetches full JIRA tickets, Confluence pages, Slack threads, or GitHub PRs/issues with all comments, reviews, linked issues, subtasks, child pages, and thread replies
Auto-detection -- pass a JIRA key (FIN-1234), Atlassian URL, Slack permalink, or GitHub PR URL and the tool figures out what to fetch
GitHub integration -- search PRs, issues, and code across repos via gh CLI; get_detail returns full PR with reviews, line comments, and CI status
Response caching -- in-memory LRU cache with configurable TTL (default 5min) eliminates redundant API calls; no_cache parameter for forced refresh
Rich Confluence output -- get_detail preserves Markdown structure (headings, tables, lists, code blocks, Confluence macros) instead of stripping to plain text
Metrics & adoption tracking -- JSONL telemetry + --stats CLI to see how often the tool is used vs individual MCPs
6 source adapters -- Slack, Confluence, JIRA, GitHub, local files (ripgrep), with a plugin architecture for adding more
Lean -- single binary, ~7MB, ~12MB RAM, 6ms startup
Preflight check -- --verify validates all credentials, scopes, and paths before first use
Why This Over Individual MCPs?
Most teams run separate JIRA, Confluence, and Slack MCP servers -- each a Node.js process with its own dependencies, memory footprint, and tool surface. To answer "what did we decide about X?", your AI assistant makes 4-5 sequential tool calls across multiple MCPs, each waiting for the previous one to finish.
unified-search-mcp replaces all of that with a single Rust binary.
Downloads a pre-built binary for your platform (macOS/Linux, x86_64/aarch64). Falls back to building from source if no binary is available. No Rust toolchain needed.
Download binary manually
Grab the latest release for your platform from Releases:
# macOS (Apple Silicon)
curl -fsSL https://github.com/ganesh-tt/unified-search-mcp/releases/latest/download/unified-search-mcp-macos-aarch64.tar.gz | tar xz
# macOS (Intel)
curl -fsSL https://github.com/ganesh-tt/unified-search-mcp/releases/latest/download/unified-search-mcp-macos-x86_64.tar.gz | tar xz
# Linux (x86_64)
curl -fsSL https://github.com/ganesh-tt/unified-search-mcp/releases/latest/download/unified-search-mcp-linux-x86_64.tar.gz | tar xz
# Linux (aarch64)
curl -fsSL https://github.com/ganesh-tt/unified-search-mcp/releases/latest/download/unified-search-mcp-linux-aarch64.tar.gz | tar xz
Build from source
git clone https://github.com/ganesh-tt/unified-search-mcp.git
cd unified-search-mcp
cargo build --release
Binary is at target/release/unified-search-mcp (~8MB). Requires Rust 1.80+.
1. Get credentials
You only need credentials for sources you want to enable. Skip any you don't use.
# Set env vars (or export them in your shell profile)
export SLACK_USER_TOKEN="xoxp-..."
export ATLASSIAN_BASE_URL="https://yourorg.atlassian.net"
export ATLASSIAN_EMAIL="[email protected]"
export ATLASSIAN_API_TOKEN="your-api-token"
# Run preflight check
./target/release/unified-search-mcp --verify --config ./config.yaml
Example output:
unified-search-mcp v0.3.0 -- preflight check
[OK] Config loaded from ./config.yaml (4 sources enabled)
[OK] Slack: auth.test OK (320ms)
[OK] Confluence: OK (180ms)
[OK] JIRA: OK (150ms)
[OK] Local text: paths accessible
/Users/you/projects/my-repo -- directory, 1247 matching files
[OK] ripgrep: ripgrep 14.1.0
Ready! 4 sources configured, 4 healthy.
4. Connect to your MCP client
Claude Code -- add to ~/.claude.json (global) or .mcp.json (per-project):
The stats command also scans Claude Code conversation logs (~/.claude/projects/) to detect when your AI assistant chose individual JIRA/Confluence/Slack MCP tools instead of unified-search. This helps measure whether unified-search is actually replacing the fragmented workflow.
Configuration Reference
All settings in config.yaml:
server:
name: "unified-search" # Server name reported to MCP clients
max_results: 20 # Global max results per query
timeout_seconds: 10 # Per-source timeout
log_level: "info" # (reserved for future use)
metrics_path: "~/.unified-search/metrics.jsonl" # Metrics log path
sources:
slack:
enabled: true/false
user_token: "xoxp-..." # Must be user token, not bot token
weight: 1.0 # Relevance multiplier (higher = ranked higher)
max_results: 10 # Max results from this source per query
confluence:
enabled: true/false
base_url: "https://..." # Your Atlassian instance URL
email: "[email protected]"
api_token: "..."
spaces: ["DEV", "OPS"] # Optional: restrict to specific spaces (empty = all)
weight: 1.0
max_results: 10
jira:
enabled: true/false
base_url: "https://..."
email: "[email protected]"
api_token: "..."
projects: ["FIN", "PLAT"] # Optional: restrict to specific projects (empty = all)
weight: 1.0
max_results: 10
github:
enabled: true/false
orgs: ["your-org"] # GitHub org(s) to search
repos: ["repo1", "repo2"] # Optional: restrict to specific repos (empty = all)
weight: 1.0
max_results: 10
local_text:
enabled: true/false
paths: # Directories to search (tilde expanded)
- "~/projects/my-repo"
include_patterns: # Glob patterns to include
- "**/*.{rs,py,js,ts,md}"
exclude_patterns: # Glob patterns to exclude
- "**/target/**"
- "**/node_modules/**"
- "**/.git/**"
weight: 0.8
max_results: 10
Environment variable interpolation: Use ${VAR_NAME} syntax in any string value. Missing env vars for disabled sources are silently ignored; missing vars for enabled sources produce a config error.
Adding a new source: Implement the SearchSource trait (4 methods: name, description, health_check, search) and register it in main.rs. See src/sources/slack.rs for a complete example.
Development
# Run tests
cargo test
# Run with verbose output
cargo test -- --nocapture
# Build debug
cargo build
# Build release
cargo build --release
# Preflight check
cargo run -- --verify --config config.yaml
# View adoption stats
cargo run -- --stats --days 7
All HTTP-based tests use wiremock for deterministic mocking.
Resource Usage
Metric
Measured value
Binary size
8MB (stripped)
Runtime RAM
~8-12MB idle
Startup time
6ms
Query latency
400ms-1.5s (parallel fan-out, bounded by slowest source)
Source lines
~7,400 Rust
Test count
207
CLI Reference
unified-search-mcp [OPTIONS]
Options:
--config <PATH> Config file path (default: config.yaml)
--verify Run preflight checks and exit
--stats Show adoption report and exit
--days <N> Days to include in stats report (default: 7, used with --stats)
Without flags, the server starts on stdio and waits for MCP JSON-RPC messages.
License
MIT
U
Unified Search
ganesh-tt·1 ดาว·
Unified search MCP server — search Slack, Confluence, JIRA, and local files in parallel